In [1]:
import sys
sys.path.append('../utils/')
In [2]:
from ImageUtils import *
In [3]:
import numpy as np
import pandas as pd # Needs the package Pandas to be installed. Check Anaconda Environments and Packages.
from sklearn.decomposition import PCA # Needs SciKit Learn package to be installed. Check Anaconda Environments and Packages.4
from sklearn.covariance import LedoitWolf
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score, precision_score
import matplotlib.pyplot as plt
from scipy.spatial.distance import mahalanobis
from collections import Counter
from sklearn.preprocessing import label_binarize
import time
from sklearn import preprocessing
import ipywidgets as widgets
import matplotlib.pyplot as plt
from scipy import stats
from sklearn.metrics import confusion_matrix
from sklearn.model_selection import KFold
from sklearn.cluster import KMeans
from sklearn.model_selection import train_test_split
from mpl_toolkits import mplot3d

Data preparation

Load dataset

In [4]:
faces94_male = readFaces94MaleFaces(gray=True)
faces94_female = readFaces94FemaleFaces(gray=True)
faces94_malestaff = readFaces94MaleStaffFaces(gray=True)

dataset = np.vstack((faces94_male, faces94_female, faces94_malestaff))

dataset_N, height, width = dataset.shape
dataset.shape
Out[4]:
(3059, 200, 180)

Data centralization and calculate of covariance matrix

In [5]:
mean_all = np.mean(dataset.reshape(dataset_N, height*width), axis=0).reshape(height, width)
plt.imshow(mean_all, plt.cm.gray)
Out[5]:
<matplotlib.image.AxesImage at 0x269fff61a90>
In [6]:
data=dataset.reshape(dataset_N, height*width) - np.mean(dataset.reshape(dataset_N, height*width), axis=0)
datasetmean=(1/(dataset_N-1))*(np.dot(data,data.T)) # Covariance matrix
print(datasetmean.shape)

(3059, 3059)

Subspaces method: Eigenfaces

Singular Value Decomposition.

In [7]:
u,s,vh = np.linalg.svd(datasetmean) # u: eigenvectors in columns; s: eigenvalues; vh = eigenvectors in rows

Face space: selection of subspace components

Option 1: Investigator's criteria of variability captured (hyperparameter)

In [8]:
representation_percentage = 0.85 # Selected variability
In [9]:
sum_eig = np.sum(s)
percentage_variance = np.divide(s, sum_eig)
sum_var = 0
num_var = 0
for i in np.arange(percentage_variance.shape[0]):
    if sum_var >= representation_percentage:
        num_var = i
        break;
    
    sum_var += percentage_variance[i]
    
num_var_select=num_var    
print("Principal components number: ",num_var_select)
print("Percent of variability captured: ",sum_var*100)
print("Images in datasets",dataset_N)

Principal components number:  48
Percent of variability captured:  85.20602884236723
Images in datasets 3059

Option 2: Investigator's criteria of threshold contribution value (hyperparameter)

In [10]:
cum_per=np.cumsum(percentage_variance)
for i in range(1,len(s)):
    change=(cum_per[i]-cum_per[i-1])/cum_per[i-1]*100
    if(change<.01):
        num_var=i-1
        print("First",num_var, "components with ",cum_per[num_var]*100,"percent of variability captured and from which the contribution is less than 0.01%")
        break

plt.figure(figsize=(12,6))
plt.plot(cum_per*100)
plt.xlabel('Number of Components')
plt.ylabel('Variance (%)') #for each component
plt.title('Cumulative Summation of the Explained Variance')
plt.show()

First 305 components with  97.58649468421936 percent of variability captured and from which the contribution is less than 0.01%

Face space

In [11]:
EigenvectorsA=np.dot(data.T,u[:,0:num_var_select])
NormEigenvectorsA = preprocessing.normalize(EigenvectorsA,axis=0, norm='l2')
print(np.linalg.norm(NormEigenvectorsA[:,5],ord=None))#check normalizacion vectores propios de XT.X

1.0000008
In [12]:
cols = 4
rows = 4
plt.figure(figsize=(30,20))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.imshow(NormEigenvectorsA[:,i].reshape(height, width), plt.cm.gray)

Projection of an image on face space

In [13]:
start=0.8
step=0.06
stop=1

facespace(percentage_variance,dataset,data,mean_all,u,dataset_N,height,width,start,step,stop,0)
In [14]:
print("Principal components number: ",num_var_select)
print("Percent of variability captured: ",sum_var*100)
print("Images in datasets",dataset_N)
print("Omega matrix facespace",np.dot(data,NormEigenvectorsA).shape)

Principal components number:  48
Percent of variability captured:  85.20602884236723
Images in datasets 3059
Omega matrix facespace (3059, 48)

specific image

In [15]:
import ipywidgets as widgets
n=widgets.BoundedFloatText(value=2690,min=0,max=dataset_N,description='image:')
display(n)
In [16]:
N_image=int(n.value)
specificimage(data,dataset,NormEigenvectorsA,mean_all,N_image,dataset_N,height,width)

random image

In [17]:
randomimage(data,dataset,NormEigenvectorsA,mean_all,dataset_N,height,width)

Distances an outliers

In [18]:
dataReconstructed=np.dot(np.dot(data,NormEigenvectorsA),NormEigenvectorsA.T)+mean_all.reshape(height*width)
print(dataReconstructed.shape)

(3059, 36000)
In [19]:
Norm=widgets.Dropdown(options=['1', '2', 'inf'],value='2',description='Norm:',disabled=False)
display(Norm)
In [20]:
if str(Norm.value)=='inf':
    ordn=np.inf
else:
    ordn=int(Norm.value)
In [21]:
edistance = np.linalg.norm(np.subtract(dataReconstructed, dataset.reshape(dataset_N, height*width)), ord=ordn, axis=1)
print(edistance.shape)

(3059,)
In [22]:
histbox(edistance)
In [23]:
threshold, outliers, zsort, indexsort, z=outlierseigenfaces(edistance,3)

print('Outliers threshold method=',np.size(outliers))
print('threshold=',threshold)
CVresult={'outliers distance':outliers,'z':zsort}
df = pd.DataFrame(CVresult)
df.sort_values('z', axis = 0, ascending = False, inplace = True, na_position ='first') 
df.head(np.size(outliers)) 

Outliers threshold method= 22
threshold= 3
Out[23]:
outliers distance z
21 5971.007812 3.918530
20 5950.271484 3.884834
19 5915.789551 3.828803
18 5882.037598 3.773957
17 5835.418457 3.698203
16 5830.108887 3.689575
15 5706.616211 3.488904
14 5696.469727 3.472417
13 5681.766602 3.448525
12 5653.831543 3.403132
11 5636.704590 3.375301
10 5627.303223 3.360024
9 5600.194336 3.315974
8 5592.637695 3.303694
7 5526.399902 3.196061
6 5495.956055 3.146591
5 5490.239258 3.137301
4 5481.522949 3.123137
3 5462.469238 3.092176
2 5453.295410 3.077269
1 5446.753906 3.066639
0 5407.920898 3.003537

low and high distance

In [24]:
fig = plt.figure(figsize=(8,10))
ax1 = fig.add_subplot(1,2,1)
plt.title("Similar Image")
ax1.imshow(dataset[indexsort[0]], plt.cm.gray)
ax2 = fig.add_subplot(1,2,2)
plt.title("Dissimilar Image")
ax2.imshow(dataset[indexsort[-1]], plt.cm.gray)
Out[24]:
<matplotlib.image.AxesImage at 0x269c01c5860>

high distances

In [25]:
cols = 4
rows = 2
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("z "+str(z[indexsort[-(i+1)]]),fontsize=20)
    plt.imshow(dataset[indexsort[-(i+1)]], plt.cm.gray)

Face recognition dataset

In [26]:
landscapes = np.array(readLandsCapeImage(gray=True))

landimages(landscapes,height,width,mean_all,NormEigenvectorsA,ordn,outliers)

No pertenece al dataset
7766.379
In [27]:
landimage=landscapes.reshape(landscapes.shape[0],height*width)-mean_all.reshape(height*width)
dataReconstructedland=np.dot(np.dot(landimage,NormEigenvectorsA),NormEigenvectorsA.T)+mean_all.reshape(height*width)
print(dataReconstructedland.shape)

edistanceland = np.linalg.norm(np.subtract(dataReconstructedland, landscapes.reshape(landscapes.shape[0], height*width)), ord=ordn, axis=1)
totaldistance=np.append(edistance,edistanceland)
histbox(totaldistance)

(210, 36000)
In [28]:
y_true=np.ones(dataset_N)
y_true=np.append(y_true,np.zeros(landscapes.shape[0]))
y_pred=(totaldistance<=outliers[0])*1

tn, fp, fn, tp = confusion_matrix(y_true, y_pred).ravel()
print('TP=', tp,'TN=',tn,'FP=',fp,'FN=', fn)
print('accuracy= ', (tp+tn)/(tp+tn+fp+fn))

plt.figure()
plt.title("Heatmap")
prediction_data = {'y_Actual': y_true,'y_Predicted': y_pred}
df = pd.DataFrame(prediction_data, columns=['y_Actual','y_Predicted'])
confusionmatrix1 = pd.crosstab(df['y_Actual'], df['y_Predicted'], rownames=['Actual'], colnames=['Predicted'])
ax=sns.heatmap(confusionmatrix1, annot=True,cmap='Blues', fmt='.0f');
ax.invert_yaxis()
ax.invert_xaxis()

TP= 3038 TN= 192 FP= 18 FN= 21
accuracy=  0.9880697460997246

false positive

In [29]:
N_land= int(np.where(edistanceland < outliers[0])[0][3])
landimage=landscapes[N_land].reshape(height*width)-mean_all.reshape(height*width)#seleccionar imagen individual
wland=np.dot(landimage,NormEigenvectorsA)#pesos w de cada Eigenface en subespacio generado
Reconstland=np.dot(wland,NormEigenvectorsA.T)+mean_all.reshape(height*width)#es mas claro w*vectores propios transpuestos
fig = plt.figure(figsize=(8,10))
ax1 = fig.add_subplot(1,2,1)
plt.title("Land image")
ax1.imshow(landscapes[N_land], plt.cm.gray)
ax2 = fig.add_subplot(1,2,2)
plt.title("Reconstructed land Image")
ax2.imshow(Reconstland.reshape(height, width), plt.cm.gray)
print('distancia',edistanceland[N_land])

distancia 4693.919
In [30]:
accuracy, tncv, fpcv, fncv, tpcv=kfold(y_true,landscapes,dataset,height,width,ordn)
CVresult={'accuracy':accuracy,'tn':tncv,'fp':fpcv,'fn':fncv,'tp':tpcv}
df = pd.DataFrame(CVresult)
df.head()   

KFold(n_splits=5, random_state=42, shuffle=True)
test:  821
faces:  611
landscapes:  210
Out[30]:
accuracy tn fp fn tp
0 0.947689 170.0 40.0 3.0 609.0
1 0.945255 166.0 44.0 1.0 611.0
2 0.950122 170.0 40.0 1.0 611.0
3 0.944039 166.0 44.0 2.0 610.0
4 0.952497 171.0 39.0 0.0 611.0

Dataset with landscapes

In [31]:
landscapes = np.array(readLandsCapeImage(gray=True))
datasetfull = np.vstack((dataset,landscapes))

datasetfull_N, height, width = datasetfull.shape
datasetfull.shape
Out[31]:
(3269, 200, 180)
In [32]:
mean_all_full = np.mean(datasetfull.reshape(datasetfull_N, height*width), axis=0).reshape(height, width)
plt.imshow(mean_all_full, plt.cm.gray)
Out[32]:
<matplotlib.image.AxesImage at 0x269bfcef278>
In [33]:
datafull=datasetfull.reshape(datasetfull_N, height*width) - np.mean(datasetfull.reshape(datasetfull_N, height*width), axis=0)
datasetmeanfull=(1/(datasetfull_N-1))*(np.dot(datafull,datafull.T))
print(datasetmeanfull.shape)

(3269, 3269)
In [34]:
u,s,vh = np.linalg.svd(datasetmeanfull)
In [35]:
representation_percentage = 0.80

sum_eig = np.sum(s)
percentage_variance = np.divide(s, sum_eig)
sum_var = 0
num_var = 0
for i in np.arange(percentage_variance.shape[0]):
    if sum_var >= representation_percentage:
        num_var = i
        break;
    
    sum_var += percentage_variance[i]
    
num_var_select=num_var    
print("Principal components number: ",num_var_select)
print("Percent of variability captured: ",sum_var*100)
print("Images in datasets",datasetfull_N)

Principal components number:  38
Percent of variability captured:  80.20745078101754
Images in datasets 3269
In [36]:
cum_per=np.cumsum(percentage_variance)
for i in range(1,len(s)):
    change=(cum_per[i]-cum_per[i-1])/cum_per[i-1]*100
    if(change<.01):
        num_var=i-1
        print("First",num_var, "components with ",cum_per[num_var]*100,"percent of variability captured and from which the contribution is less than 0.01%")
        break

plt.figure(figsize=(12,6))
plt.plot(cum_per*100)
plt.xlabel('Number of Components')
plt.ylabel('Variance (%)') #for each component
plt.title('Cumulative Summation of the Explained Variance')
plt.show()

First 415 components with  97.31403589248657 percent of variability captured and from which the contribution is less than 0.01%
In [37]:
EigenvectorsA=np.dot(datafull.T,u[:,0:num_var_select])
NormEigenvectorsA = preprocessing.normalize(EigenvectorsA,axis=0, norm='l2')
print(np.linalg.norm(NormEigenvectorsA[:,5],ord=None))#check normalizacion vectores propios de XT.X

1.0000013
In [38]:
start=0.8
step=0.06
stop=1

facespace(percentage_variance,datasetfull,datafull,mean_all_full,u,datasetfull_N,height,width,start,step,stop,dataset_N)
In [39]:
print("Principal components number: ",num_var_select)
print("Percent of variability captured: ",sum_var*100)
print("Images in datasets",datasetfull_N)
print("Omega matrix facespace",np.dot(datafull,NormEigenvectorsA).shape)

Principal components number:  38
Percent of variability captured:  80.20745078101754
Images in datasets 3269
Omega matrix facespace (3269, 38)

random image

In [40]:
N_image= np.random.randint(3059, high=datasetfull.shape[0], size=1)[0]
Image=datafull[N_image]
w=np.dot(Image,NormEigenvectorsA)
Reconstructed=np.dot(w,NormEigenvectorsA.T)+mean_all_full.reshape(height*width)
example_image = Reconstructed
original_image = datasetfull[N_image]
fig = plt.figure(figsize=(8,10))
ax1 = fig.add_subplot(1,2,1)
plt.title("Original Image")
ax1.imshow(original_image, plt.cm.gray)
ax2 = fig.add_subplot(1,2,2)
plt.title("Reconstructed Image")
ax2.imshow(example_image.reshape(height,width), plt.cm.gray)
Out[40]:
<matplotlib.image.AxesImage at 0x269bfc67f98>
In [41]:
datafullReconstructed=np.dot(np.dot(datafull,NormEigenvectorsA),NormEigenvectorsA.T)+mean_all_full.reshape(height*width)
print(datafullReconstructed.shape)
edistancefull = np.linalg.norm(np.subtract(datafullReconstructed, datasetfull.reshape(datasetfull_N, height*width)), ord=ordn, axis=1)
print(edistancefull.shape)
histbox(edistancefull)

(3269, 36000)
(3269,)
In [42]:
z = np.abs(stats.zscore(edistancefull))
percentil_93, percentil_94 = np.percentile(edistancefull, [93, 94])
outliersindex=np.where(edistancefull > percentil_93)
outliersfull=edistancefull[outliersindex]
zsort=z[outliersindex]
indexsortout=np.argsort(outliersfull)
outliersfull=outliersfull[indexsortout]
zsort=zsort[indexsortout]

indexsort=np.argsort(edistancefull) #indice de imagenes distance de menor a mayor
edistancesort=edistancefull[indexsort] #distancias de imagenes menor a mayor

print('Outliers threshold method=',np.size(outliersfull))
print('percentil_93=',percentil_93)
#print('Q1= ',quartile_1, 'Q3= ',quartile_3)
CVresult={'outliers distancefull':outliersfull,'z':zsort}
df = pd.DataFrame(CVresult)
df.sort_values('z', axis = 0, ascending = False, inplace = True, na_position ='first') 
df.head(10) 

Outliers threshold method= 229
percentil_93= 5338.332343750001
Out[42]:
outliers distancefull z
228 12027.106445 6.893754
227 11252.820312 6.218557
226 10952.622070 5.956776
225 10923.311523 5.931216
224 10741.379883 5.772567
223 10731.321289 5.763796
222 10700.211914 5.736668
221 10642.039062 5.685939
220 10499.667969 5.561788
219 10481.004883 5.545513
In [43]:
cols = 4
rows = 2
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("z "+str(z[indexsort[-(i+1)]]),fontsize=20)
    plt.imshow(datasetfull[indexsort[-(i+1)]], plt.cm.gray)
In [44]:
y_true=np.ones(dataset_N)
y_true=np.append(y_true,np.zeros(landscapes.shape[0]))
y_pred=(edistancefull<=percentil_93)*1
#y_pred=(z <= threshold)*1

tn, fp, fn, tp = confusion_matrix(y_true, y_pred).ravel()
print('TP=', tp,'TN=',tn,'FP=',fp,'FN=', fn)
print('accuracy= ', (tp+tn)/(tp+tn+fp+fn))

plt.figure()
plt.title("Heatmap")
prediction_data = {'y_Actual': y_true,'y_Predicted': y_pred}
df = pd.DataFrame(prediction_data, columns=['y_Actual','y_Predicted'])
confusionmatrix1 = pd.crosstab(df['y_Actual'], df['y_Predicted'], rownames=['Actual'], colnames=['Predicted'])
ax=sns.heatmap(confusionmatrix1, annot=True,cmap='Blues', fmt='.0f');
ax.invert_yaxis()
ax.invert_xaxis()

TP= 3013 TN= 183 FP= 27 FN= 46
accuracy=  0.9776690119302539

Part 3 Classification

Unsupervised image classification - K means

In [45]:
print("Principal components number: ",num_var_select)
print("Percent of variability captured: ",sum_var*100)
print("Images in datasets",datasetfull_N)
print("Omega matrix facespace",np.dot(datafull,NormEigenvectorsA).shape)

Principal components number:  38
Percent of variability captured:  80.20745078101754
Images in datasets 3269
Omega matrix facespace (3269, 38)
In [46]:
omegaw=np.dot(datafull,NormEigenvectorsA)
print(omegaw.shape)

(3269, 38)
In [47]:
kmeans = KMeans(n_clusters=3, random_state=42).fit(omegaw)
wcentroids=kmeans.cluster_centers_
wcentroids.shape
Out[47]:
(3, 38)
In [48]:
cols = 3
rows = 1
plt.figure(figsize=(12,8))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("Class "+str(i+1))
    plt.imshow((np.dot(kmeans.cluster_centers_[i],NormEigenvectorsA.T)+mean_all_full.reshape(height*width)).reshape(height, width), plt.cm.gray)
In [49]:
y_label=kmeans.labels_
wtotaldist=kmeans.transform(omegaw)
wdistances = np.amin(wtotaldist, axis=1)

print(wdistances.shape[0])

3269

Class 1

In [50]:
kclass=0
print("Number images: "+str(wdistances[y_label==kclass].shape[0]))

Number images: 1084
In [51]:
histbox(wdistances[y_label==kclass])
In [52]:
CVresult={'w distances':wdistances,'label':y_label}
df = pd.DataFrame(CVresult)
df.sort_values('w distances', axis = 0, ascending = True, inplace = True, na_position ='first')
df2=df.loc[df[df.columns[1]]==kclass]
df2.head(6)
Out[52]:
w distances label
3034 4826.396973 0
3033 5063.331543 0
3254 5128.105957 0
3024 5162.825684 0
3029 5168.416992 0
3031 5169.007812 0
In [53]:
#low distances
cols = 4
rows = 1
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("Class1 low distance "+ str(df2['w distances'][df2.index[i]]),fontsize=20)
    plt.imshow(datasetfull[df2.index[i]], plt.cm.gray)
In [54]:
#High distances
cols = 4
rows = 1
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("Class1 high distance "+ str(df2['w distances'][df2.index[-(i+1)]]),fontsize=20)
    plt.imshow(datasetfull[df2.index[-(i+1)]], plt.cm.gray)

Class 2

In [55]:
kclass=1
print("Number images: "+str(wdistances[y_label==kclass].shape[0]))

Number images: 1174
In [56]:
histbox(wdistances[y_label==kclass])
In [57]:
CVresult={'w distances':wdistances,'label':y_label}
df = pd.DataFrame(CVresult)
df.sort_values('w distances', axis = 0, ascending = True, inplace = True, na_position ='first')
df2=df.loc[df[df.columns[1]]==kclass]
df2.head(6)
Out[57]:
w distances label
382 3412.193359 1
399 3432.506348 1
397 3450.682373 1
383 3455.498291 1
1672 3465.895508 1
1670 3483.517090 1
In [58]:
#low distances
cols = 4
rows = 1
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("Class2 low distance "+ str(df2['w distances'][df2.index[i]]),fontsize=20)
    plt.imshow(datasetfull[df2.index[i]], plt.cm.gray)
In [59]:
#High distances
cols = 4
rows = 1
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("Class2 high distance "+ str(df2['w distances'][df2.index[-(i+1)]]),fontsize=20)
    plt.imshow(datasetfull[df2.index[-(i+1)]], plt.cm.gray)

Class 3

In [60]:
kclass=2
print("Number images: "+str(wdistances[y_label==kclass].shape[0]))

Number images: 1011
In [61]:
histbox(wdistances[y_label==kclass])
In [62]:
CVresult={'w distances':wdistances,'label':y_label}
df = pd.DataFrame(CVresult)
df.sort_values('w distances', axis = 0, ascending = True, inplace = True, na_position ='first')
df2=df.loc[df[df.columns[1]]==kclass]
df2.head(6)
Out[62]:
w distances label
1546 4559.927734 2
1542 4578.041504 2
1548 4609.369629 2
340 4704.488281 2
1547 4705.103516 2
1557 4708.507812 2
In [63]:
#low distances
cols = 4
rows = 1
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("Class3 low distance "+ str(df2['w distances'][df2.index[i]]),fontsize=20)
    plt.imshow(datasetfull[df2.index[i]], plt.cm.gray)
In [64]:
#High distances
cols = 4
rows = 1
plt.figure(figsize=(25,15))
for i in np.arange(rows * cols):
    plt.subplot(rows, cols, i + 1)
    plt.title("Class3 high distance "+ str(df2['w distances'][df2.index[-(i+1)]]),fontsize=20)
    plt.imshow(datasetfull[df2.index[-(i+1)]], plt.cm.gray)

Results K-means

In [65]:
y_true=np.ones(faces94_male.shape[0])*2
y_true=np.append(y_true,np.ones(faces94_female.shape[0]))
y_true=np.append(y_true,np.ones(faces94_malestaff.shape[0])*2)
y_true=np.append(y_true,np.zeros(landscapes.shape[0]))
In [66]:
cm=confusion_matrix(y_true, y_label).ravel()

plt.figure()
plt.title("Heatmap")
prediction_data = {'y_Actual': y_true,'y_Predicted': y_label}
df = pd.DataFrame(prediction_data, columns=['y_Actual','y_Predicted'])
confusionmatrix1 = pd.crosstab(df['y_Actual'], df['y_Predicted'], rownames=['Actual'], colnames=['Predicted'])
ax=sns.heatmap(confusionmatrix1, annot=True,cmap='Blues', fmt='.0f');
ax.xaxis.set_ticklabels(['landscape', 'female','male']); ax.yaxis.set_ticklabels(['landscape', 'female','male']);
ax.invert_yaxis()
In [67]:
print('accuracy= ', (cm[0]+cm[4]+cm[8])/np.sum(cm))
print('accuracy= ', cm[0]/y_true[y_true==0].shape[0],cm[4]/y_true[y_true==1].shape[0],cm[8]/y_true[y_true==2].shape[0])

accuracy=  0.40532272866319974
accuracy=  0.5619047619047619 0.656641604010025 0.35526315789473684

Demo K-means

In [68]:
kmeansk, NormEigenvectorsAk, X_train, X_test, y_predk= kmeansplit(datasetfull,y_label,datasetfull_N,height,width,representation_percentage,3)
In [69]:
print("Number images Test: "+str(y_predk.shape[0]))
print("Number images class 1: "+str(y_predk[y_predk==0].shape[0]))
print("Number images class 2: "+str(y_predk[y_predk==1].shape[0]))
print("Number images class 3: "+str(y_predk[y_predk==2].shape[0]))

Number images Test: 654
Number images class 1: 235
Number images class 2: 208
Number images class 3: 211
In [70]:
N_image= np.random.randint(0, high=X_test.shape[0], size=1)[0]
print('Demo: predict new image as class '+ str(y_predk[N_image]+1))
Image=X_test[N_image]#seleccionar imagen individual
fig = plt.figure(figsize=(8,10))
ax1 = fig.add_subplot(1,2,1)
plt.title("Image")
ax1.imshow(Image, plt.cm.gray)
ax2 = fig.add_subplot(1,2,2)
plt.title("class "+str(y_predk[N_image]+1))
ax2.imshow((np.dot(kmeansk.cluster_centers_[y_predk[N_image]],NormEigenvectorsAk.T)+np.mean(X_train.reshape(X_train.shape[0], height*width), axis=0)).reshape(height, width), plt.cm.gray)

Demo: predict new image as class 3
Out[70]:
<matplotlib.image.AxesImage at 0x269c0341f60>

Demo female

In [71]:
dataclass1=faces94_female
dataset_N2, height, width = dataclass1.shape
mean_class1 = np.mean(dataclass1.reshape(dataset_N2, height*width), axis=0)
plt.imshow(mean_class1.reshape(height, width), plt.cm.gray)
Out[71]:
<matplotlib.image.AxesImage at 0x269bed01240>
In [72]:
omegaclass1=mean_class1 - np.mean(dataset.reshape(dataset_N, height*width), axis=0)
wclass1=np.dot(omegaclass1,NormEigenvectorsA)#pesos w de cada Eigenface en subespacio generado
Reconstructedc1=np.dot(wclass1,NormEigenvectorsA.T)+mean_all.reshape(height*width)#es mas claro w*vectores propios transpuestos


fig = plt.figure(figsize=(8,10))
ax1 = fig.add_subplot(1,2,1)
plt.title("Female mean")
ax1.imshow(mean_class1.reshape(height, width), plt.cm.gray)
ax2 = fig.add_subplot(1,2,2)
plt.title("Reconstructed Female mean")
ax2.imshow(Reconstructedc1.reshape(height, width), plt.cm.gray)
Out[72]:
<matplotlib.image.AxesImage at 0x269bfba16a0>
In [73]:
class1=dataclass1.reshape(dataset_N2, height*width)-mean_all.reshape(height*width)
wclass1total=np.dot(class1,NormEigenvectorsA)
print(wclass1total.shape)
print(wclass1.shape)

(399, 38)
(38,)
In [74]:
Norm=2
edistanceclass1 = np.linalg.norm(np.subtract(wclass1total, wclass1), ord=Norm, axis=1)
print(edistanceclass1.shape)
print(wclass1.shape)

(399,)
(38,)
In [75]:
import matplotlib.pyplot as plt

plt.figure(figsize=(15,4))
plt.subplot(1,2,1)
plt.title('Histogram')
plt.grid(True)
plt.hist(edistanceclass1);
plt.subplot(1,2,2)
plt.title('Boxplot')
plt.boxplot(edistanceclass1, 0, 'rs', 0);
plt.show()
In [76]:
quartile_1, quartile_3 = np.percentile(edistanceclass1, [25, 75])


image_index = np.random.randint(0, high=dataset_N, size=1)[0]
Imagetest=dataset[image_index].reshape(height*width) - np.mean(dataset.reshape(dataset_N, height*width), axis=0)
wtest=np.dot(Imagetest,NormEigenvectorsA)

Norm=2
test = np.linalg.norm(np.subtract(wtest, wclass1), ord=Norm, axis=0)

if test>quartile_3:
        print('no es mujer')
else:
        print('es mujer')
print(quartile_3) 
print(test)    
   
plt.imshow(dataset[image_index], plt.cm.gray)

es mujer
8560.126953125
7001.064
Out[76]:
<matplotlib.image.AxesImage at 0x269b8115ba8>
In [77]:
Norm=2
edistanceclass1t = np.linalg.norm(np.subtract(np.dot(data,NormEigenvectorsA), wclass1), ord=Norm, axis=1)
print(edistanceclass1t.shape)
print(wclass1.shape)
import matplotlib.pyplot as plt

plt.figure(figsize=(15,4))
plt.subplot(1,2,1)
plt.title('Histogram')
plt.grid(True)
plt.hist(edistanceclass1t);
plt.subplot(1,2,2)
plt.title('Boxplot')
plt.boxplot(edistanceclass1t, 0, 'rs', 0);
plt.show()

(3059,)
(38,)